Research On Face Anti-spoofing Detection For Photo And Video Playback Attacks

With the rapid development of artificial intelligence,as an important field in computer vision,face recognition is widely used in online payment,security check,unmanned supermarket,access control,mobile phone unlocking and other identity authentication scenarios.At present,face recognition systems still have the security problem of being attacked by faking face,and face anti-spoofing detection is a technology proposed in recent years to solve this problem,which aims to promote the development of face recognition technology.In this thesis,based on summarizing the literature at home and abroad,face anti-spoofing detection based on traditional artificial designed features and adaptive scheme based on deep learning for the photo and video playback attack problem are studied.The main work of this thesis is as follows.(1)Several texture features are extracted in different colour spaces to compare the impact of multiple colour features and texture features on face anti-spoofing detection.The face image is divided into patches,and the patch features of face images with multiple local pattern operators are extracted in different colour spaces.Finally,the SVM classifier is used to achieve the discrimination between actual faces and fake faces,and the experimental results on three public datasets show the effectiveness of the patch features.(2)A differential quantized adjacent local binary pattern(DQ?Co ALBP)descriptor is proposed to synthesize the difference between local centroids and surrounding points of images in different directions,and fuse with LPQ features in different colour spaces.Finally,we use the SVM classifier to realize the discrimination between actual faces and fake faces.The experimental results show that the DQ?Co ALBP descriptor is effective in face anti-spoofing detection.The experimental results also demonstrate that the fusion of DQ?Co ALBP and LPQ descriptor features can achieve higher recognition performance.(3)To address the domain adaptive problem of face anti-spoofing detection algorithms,a domain adaptive scheme based on ADDA-ResNet network is proposed to extract deep convolutional features,inspired by GAN network,alternately optimizing the domain discriminator and feature encoder,adjusting the parameters of the target domain feature encoder,finding the common feature space distribution of the source domain encoder to reduce the difference of feature distribution between the target domain and the source domain.Compared with the experimental results of other methods,it has been verified that the ADDA-ResNet model has higher adaptive detection ability in the unknown domain.